基于SVM和有监督描述子学习算法的脑MR图像颅骨分割方法

Automated Segmentation Based on Support Vector Machine and Supervised Descriptor Learning from Brain MR Image

神经电流源定位研究首先要解决EEG、MEG正问题的计算。在求解MEG和EEG正问题的过程中,为了精确地计算传导矩阵,常常需要对脑组织进行分层建模。在脑MR图像中,虽然软组织能被清晰地成像,但颅骨却由于缺少氢而呈现低灰度值,从而很难自动分割出颅骨。因此如何从脑MR图像中准确、自动分割出颅骨是解决MEG、EEG正问题的关键。为解决上述问题,提出一种基于支持向量机的自动脑MR图像颅骨分割方法,提取病人MR图像的全局特征和局部特征进行训练,并结合有监督描述子学习算法SDL,将得到的特征矩阵进行压缩,去掉冗余的特征,得到一个紧凑的特征描述,最终利用SVM从脑MR图像中自动识别出骨骼。实验结果表明,采用支持向量机结合有监督描述子学习算法的分割方法与仅使用支持向量机和仅使用数学形态学方法相比,分割精度进一步提升,Dice分割精度分别为0.832,0.798,0.482,从而解决了从脑MR图像自动分割颅骨的任务,并为解决EEG和MEG正问题的研究奠定基础。

A solution of EEG/MEG forward problem is essential and important in stereotactic neurosurgery applications. It is necessary to bui ld a multi-layer brain model to distinguish different tissues for MEG/EEG forward problem. Although soft tissues can be clearly seen in MR images, but the intensity of skull is so low because of a lack of hydrogen in skull that can？t be segmented auto-matically and accurately from MR image. Extracting skull form MR image automatically end up to be a key problem when calculating the MEG/EEG forward problem. In order to solve the above problem, a support vector machine（SVM） is proposed based segmentation algorithm using global features and local features of MR image. Moreover, the supervised descriptor learning （SDL） algorithm is com-bined that can transform the feature matrix into a compact one, and finally the skull from brain MR image is extrated by training on multi-modal images from the same patient whose CTs and MRs are available. Compared to the algorithm based on SVM only and math-ematical morphology based algorithm, the proposed method shows a considerable improvement on segmentation accuracy. The pro-posed method achieves an accuracy with Dice coefficient 0.832 compared with the other two methods 0.798 and 0.482. The proposed hybrid algorithm extract the skull successfully, so that the EEG,MEG source imaging problem can be solved easily in future work.